* Revert for testing TensorRT 7.1
* change to origianl googletest version
* change machine
* remove build arg
* change back machine
* revert back googletest version
* Make it ready to merge to master
* revert onnx-tensorrt to v7.1
* rename yml
* use [[ ]] in bash command
* add sudo
* add chmod
* add correct path
* change another way to revert onnx-tensorrt
* change docker image to manylinux build
* Added support for ReduceMean on DNNL EP for CPU and GPU
Signed-off-by: Chethan Palangotu Keshava <chethan.palangotu.keshava@intel.com>
* Added fix for a resnet model failure where it was failing to create dst shape for reducemean when it was part of a subgraph with other ops
Signed-off-by: Chethan Palangotu Keshava <chethan.palangotu.keshava@intel.com>
* Removing the DNNL EP from these unit tests. This is in anticipation of two changes:
- DNNL EP unit tests would be added in a different location later on, so addition of EP individually to these tests will not be necessary
- This was causing a memory leak fail in debug build. The bug is in the EP itself and not in the code added for reducemean. The fix for this is in the i/o handling overhaul which will be added later.
* Update reduction_ops_test.cc
Had accidentally deleted a new line. Making sure there are no unnecessary changes in this file
**Description**:
Enforce no repetition of n-grams. Scores are set to `-inf` for tokens that form a repeated n-gram if added to the back of the input_ids.
**Motivation and Context**
Needed by transformer models in sequence generation algorithms (greedy search and beam search). This module has heavy impact on performance, and can be highly parallelized.
* clean up builds for interop_torch
* add python dependency for executables
* disable onnxruntime_ENABLE_TRAINING_TORCH_INTEROP by default; enable it in ortmodule GPU training pipeline only
* disable training unrelated tests when torch interop is enabled
* simplify the python dependency.
* clean up and fix
- Allow anyone to kick off a perf test here. Customize: branch, eps, model selection, cuda version.
- Only run shape inference when required.
- Kill errored out memory processes.
- Remove warmup run.
- Clean up script.
- Standalone_TRT is it's own "EP" vs as an additional run with TRT EP
* Enable slice for NNAPI EP
* Add ANEURALNETWORKS_STRIDED_SLICE support
* Addressed CR comments
* Addressed CR comments, rename PrepareForCompute to PrepareForComputeHelper to avoid confusion
* Create a shared quantization util for all unit tests.
* Cleanup qlinear_binary_op_test.cc
* save
* save
* save
* cleanup
* save
* cleanup for linux build
* Fix PythonOp with input has no gradient
* Fix another bug which happens when inputs require gradient
* Remove comments
Co-authored-by: Peng Wang <pengwa@microsoft.com>
* Change full ort to mobile ort
* Update Android example to load mobile ort
* Change the format of test models to ort
* update ios to use mobile ort
* revise README
* use onnxruntime-mobile-c CocoaPods in a npm package
Check the inputs of all nodes are part of the subgraph for all
operators. Previously the code assumed all operators only had
a single input except for the "Sum" operator.
This resolves issue seen when adding new operators that a subgraph
was incorrectly accepting a node when the subgraph should not have
because it was not following the topology of the nodes.
Signed-off-by: George Nash <george.nash@intel.com>
* Avoid hashing the operator type in the GraphViewer priority node check unless the string has a chance of matching.
Below are perf numbers from a test that loads 16 models multiple times. I was checking that some unrelated changes didn't have unexpected perf cost and found the PriorityNodeCompare overwhelmed any contribution the other changes were making.
*Before*
CPU Time:74.678s
CPU Time for relevant Top Hotspots
std::_Hash_array_representation<char> 20.834s
onnxruntime::PriorityNodeCompare::IsHighPri 7.589s
onnxruntime::Graph::KahnsTopologicalSort 4.487s
*After*
CPU Time:47.103s
CPU Time for relevant Top Hotspots
onnxruntime::Graph::KahnsTopologicalSort 4.465s
onnxruntime::PriorityNodeCompare::IsHighPri 2.873s
Adding support for generating API documentation with the Jazzy tool.
It's a manual process now, but we can eventually make it a part of the release pipeline.
* Merged PR 6093117: Fix test_DynamicQuantizedLinear_max_adjusted_expanded by allowing Identity operator to run on non-float inputs
Motivation:
As part of the OnnxConformance Backend tests, DynamicQuantizedLinear_max_adjusted_expanded is failing.
Root Cause:
- The test model has `Identity` operator as one of the node. The input of this node is of non-float data type.
- In DML, `Identity` operator is registered as operator which requires floating input.
- As per `DirectMLSchema.h`, support for non-float input has been added for `Identity` operator in DML but the same has not been reflected in the `OperatorRegistration.cpp`.
Changes:
- Removed all traces of the requiresFloatFormatsForGraph flag from it's definition and usage. This flag was only used for Identity and it's related operator.
- Added null check for the graphOutput nodeArg in GraphDescBuilder.cpp to stop the crash of the test.
Related work items: #33076298
* Merged PR 6103324: Remove usage of non-generic error code (FWP_E_NULL_POINTER)
Motivation:
Addressing Dwayne comment on the previous PR. [Ref: [6093117](https://dev.azure.com/microsoft/WindowsAI/_git/onnxruntime/pullrequest/6093117?discussionId=44292162&path=%2Fonnxruntime%2Fcore%2Fproviders%2Fdml%2FDmlExecutionProvider%2Fsrc%2FGraphPartitioner.cpp)]
Changes:
Inside the DML EP, we should not use some other platform specific error codes. Instead we should a appropriate generic error code.
Related work items: #33076298
Co-authored-by: Sumit Agarwal <sumitagarwal@microsoft.com>
* Add enhanced partitioning utils and convert internal testing EP to use it. Will convert NNAPI EP once checked in.
Background:
Currently most EPs do their partitioning by iterating the model in the topologically sorted order. Whilst this works, it doesn’t ensure that all nodes which could possibly be added to the current group are, as the group is closed as when the first unsupported node is seen.
Changes:
- Ask EP for all nodes it supports first
- Do partitioning aware topological sort
- Groups nodes and flips between processing supported and unsupported nodes to maximise inputs that will be available for each partition
- Create groups of nodes for the partition using the new order of nodes
- Create ComputeCapability for each group
There’s also an additional ability to specify operators to stop at. The processing will find all downstream nodes from ‘stop at’ operators and exclude them. If NonMaxSuppression is specified we can prevent the post-processing from SSD Mobilenet and MobileDet attempting to use NNAPI (so easy way to have parity with the TF Lite behavior). I don’t think there’s an automated way to determine what if any ‘stop at’ operators are required for a model, so this will need to be a configuration parameter for the EP and we’ll need to document recommended values for popular models.
* Address #7981
Reworked the logic around forcing a node to run on fp32 even if it was supported on fp16.
The github issue had multiple factors. In ORT 1.8 we remove Identity nodes that produce graph outputs as they're not needed. That resulted in a Loop node no longer having output nodes (it produces graph outputs instead), which meant the check in IsSingleInputNodeFloat16Node returned true as there was no longer a downstream Identity node processing fp16 data.
We shouldn't only force a node to fp32 in very specific circumstances, and the changes hopefully check for those more precisely.
